HFENet: High-Frequency Enhanced Network for Shape-Aware Segmentation of Left Ventricle in Pediatric Echocardiograms

Tianxiang Chen, Ziyang Wang, Zi Ye

Abstract: Automated ventricular function analysis can make healthcare more consistent and available, especially where resources are scarce. However, current segmentation methods trained on adult heart ultrasounds cannot finely delineate the irregular shape of the left ventricle due to the ignorance of boundary feature exploration. To address this challenge, we introduce HFENet for shape-aware left ventricle segmentation. We propose a High-Frequency Enhancement Block (HFEB) that focuses on enhancing the high-frequency component, which is also the boundary area of left ventricles in pediatric echocardiograms. This way, the target boundary details can be explored during feature extraction. We propose space-frequency consistency loss to refine the shape of predicted masks further. Specifically, our new loss function incorporates spatial and frequency domain loss components to jointly refine predicted mask shapes in cases where current spatial-domain segmentation losses cannot be optimized further. Experiments carried out on two public datasets prove the superiority of the proposed HFENet in predicting the fineness of target shapes.